AI and the future of work: where the optimists, the skeptics, and the data converge.
An economist who runs experiments (Alex Imas, UChicago Booth), a founder who has automated his own company as far as it will go (Dan Shipper, Every), a Wall Street chief executive (David Solomon, Goldman Sachs), and the Stanford economist who measures the data (Erik Brynjolfsson) all circle the same question: what happens to human work as the machines get good? Three of them arrive, from very different directions, at more expert human work, not less. Brynjolfsson shares much of that optimism — and then insists that the gains and the disruption are two separate questions with two separate answers. Here is where the four converge, and where they pull apart — and how the field votes when The Wall Street Journal puts the same question to sixteen economists, Nobel laureate included.
Six lenses + one ballot · five sources
01The history is louder than the headlines
The assumption that the set of jobs we can see today is the set of jobs that will ever exist. If you believe that, automation looks like a one-way march to zero employment. Two centuries of evidence — and one Every-sized natural experiment happening right now — say otherwise.
In 1820, David Ricardo wrote On Machinery and changed his mind about the Industrial Revolution: technology, he warned, could put people permanently out of work. Two hundred years later, U.S. prime-age employment is at one of its highest readings ever. The jobs of 1820 are mostly gone. The number of jobs is not.
Why? As one sector becomes cheap to produce — agriculture, manufacturing — its share of GDP shrinks. People get richer. Dollars and time flow toward desires that couldn't be satisfied before: pet care, private tutoring, therapists, personal trainers, Michelin restaurants. Most of these categories did not exist as jobs in 1940.
"At Every, we've automated everything we can. We use Codex and Claude Code across coding, writing, design, customer service, and more… And yet it seems like, for us, there's more human work to do than ever. We are a team of almost 30 people, and we haven't fired all of our employees in favor of agents."
"In short, the future looks weird, but also familiar."
"The United States has a long track record of creating new jobs in response to disruption, from the electrification of the 1900s to the digital revolution of the 1990s; I don't see any reason to think this dynamic will stop now."
The receipts he points at: since 1962, U.S. civilian employment rose ~145% while the working-age population rose ~128%. Manufacturing employment fell from 15.5M → 12.5M over the same span, but health care — barely a sector in 1962 — now employs more than 18 million Americans. The mix shifts. The total grows.
If Ricardo were right — vs. what actually happened
Where they converge
Imas reads the macro record. Shipper reads his own 30-person org. Solomon reads Goldman's headcount over six decades. All three land on the same anomaly: the doom forecast keeps being wrong, every time, at every scale.
02Cheaper makes bigger
Make a thing more efficient, and demand can rise, not fall — because the cheaper version unlocks uses that weren't viable before. Named for William Stanley Jevons, who in 1865 noticed it with coal. In 2026 you can watch it happen inside any GitHub repository.
Each engine burned less coal — yet total coal use tripled
This isn't ordinary supply-and-demand. It's the paradox that efficiency itself expands the market. Imas points at software as today's coal:
If one AI-augmented engineer can do the work of ten, naively you'd fire nine. But cheaper software means new buyers — small firms, niche use cases, individuals — who could never afford it before. If demand is elastic enough, the firm hires more, not fewer.
The spreadsheet is the perfect prior analog. In the 1970s, Gunnar Myrdal wrote to the White House warning that digitized spreadsheets would wipe out accountants. Instead, the number of accountants quadrupled.
"When the cost goes down for something previously rare, supply suddenly goes way up. At Every, we see this all the time. Operations and customer service people are writing code and issuing pull requests. Marketers are making YouTube thumbnails. Engineers and product people are writing drafts of articles, guides, and landing pages when they never would have before."
"Take OpenClaw, the open-source AI-agent project: as of May 16, 2026, its repository had already seen 44,469 pull requests, including 12,430 since April 1 and 3,990 since May 1. For comparison, Kubernetes — one of the most popular open-source projects in the world — got 5,200 pull requests in all of 2022."
The microfiche-to-Excel story. "As a first-year banking analyst, making a graph of a stock's performance took me six hours of looking up prices on WSJ microfiche." Today that graph takes seconds. Goldman didn't fire the analysts — "we've employed more people than ever in recent years." The tool raised the bar on what each analyst produces; the headcount went up, not down.
The ATM story. ATMs were the canonical "this technology eliminates jobs" prediction of the 1970s and 80s. Forty years on, with ATMs, digital banking, and decades of consolidation, U.S. commercial banking employment is roughly the same as the mid-1990s. Branches got cheaper, so banks opened more of them. The teller's role evolved; it didn't vanish.
The data-center build-out. The same wave automating away white-collar tasks created more than 200,000 construction jobs since 2022 to build the A.I. infrastructure itself. Goldman economists' figure. The demand the technology unlocks is showing up in adjacent sectors.
Jevons isn't automatic — it depends on the demand curve. When AI makes a worker 10× more productive, you hire 10× the workers only if demand for the output is elastic. Brynjolfsson's two reference cases: agriculture is inelastic — farms got vastly more productive and employment fell from ~90% of Americans to ~2%, because we don't eat ten times more food. Commercial aviation is elastic — jets made each pilot far more productive, flying got cheap, demand exploded, and pilot employment grew.
The same split is now running through software itself: entry-level coding, mostly codified and repeatable, is contracting; senior engineering, where judgment creates new value, sees more demand and higher pay. Which side a job lands on, he argues, gets sorted by decentralized entrepreneurship, not by forecasting — "this is the job of entrepreneurs, and this is why you have a decentralized economy."
His advice to workers is unsentimental: "Use these tools, become as efficient as you can. It may increase demand, and it may increase wages — or it may not. But if you don't, somebody else is going to do it and outcompete you."
Where they converge
Same mechanism, three vantage points — plus a caveat. Imas observes Jevons in BLS hiring data. Shipper observes it inside Slack and Git. Solomon points at his own résumé and four decades of commercial banking. Brynjolfsson adds the switch: the multiplier only runs forward when demand is elastic — and for some jobs it isn't.
03Tasks unbundle. Jobs don't.
Most jobs are bundles of interdependent tasks. Automate four of five and the fifth still needs a human — or the whole thing collapses, like the cold-stiffened O-ring that destroyed the Challenger. Shipper has a complementary frame: every agent at Every sits sandwiched between humans on both ends. Pull either end away and the agent stops being useful.
The economist's old model treats tasks as independent: automate one, the others sit fine. The O-ring model (Goldfarb & Gans) says no — over-salt a Michelin meal and the rest of the perfection doesn't save it. Most knowledge work looks like this.
A radiologist's job is a strong bundle: reading the scan, explaining it to a patient, coordinating the care team, querying the EHR, escalating to a surgeon. The tasks need each other. You can't peel one out without breaking the rest.
A long-haul trucker is a weak bundle (Garicano): driving, safety checks, paperwork, dock handoff. Each piece can be automated separately. Different exposure.
A subtler implication: in strong bundles, the person left behind may earn more, not less — they now have the time and the supporting AI to do the irreplaceable part really well. Wages can rise even as headcount falls. Whether headcount actually falls depends on whether demand for the bundled output is elastic (back to Jevons).
Strong bundle vs. weak bundle
"Across both forms — coworker agents and embedded agents — the pattern is the same. Employee agents take over more of the stable, repeatable, well-framed layer of work. But there is a lot of work that still requires a human being in the loop." Kieran Klaassen, GM at Every's Cora, calls it the human sandwich:
"The further away an agent gets from a human who is in charge of making sure it works well, the less well it works. In our initial internal roll-out of agents as employees, we gave every employee an agent, but we soon moved back to agents that work for a particular team or the whole company. Agents need a lot of maintenance, and personal agents quickly get stale when the employees they were working with gave up on them."
"While A.I. eliminates jobs in some sectors, it may lead to job growth in others. Goldman Sachs may need fewer people in regulatory reporting or client onboarding, freeing us up to hire more bankers, traders and asset managers."
That's the bundle in action. Banking is a strong bundle: client trust, judgment under uncertainty, structuring a deal, sitting across from a CEO. Regulatory reporting and onboarding are the weak-bundle tasks at the edges. Peel those off and the high-judgment core gets more time, more capacity, and more hires.
"'Can be replaced' does not mean 'will be replaced.'" Solomon's example: despite ATMs, digital banking, and decades of consolidation, commercial banking employment is roughly flat to the mid-1990s. The relational, judgment-laden parts of the job were always the real value — automation reorganized around them.
Why does so much AI get built to replace the human in the bundle rather than complement them? Brynjolfsson blames a 75-year-old default. By defining the goal as imitating humans, the Turing Test quietly pointed the whole field at substitution. "I think Alan Turing… sent us in the wrong direction when he said the test of AI is how close it can mimic humans. Instead of trying to have machines do things that are hard for machines and easy for humans, we should work on having machines do things that are easy for machines and hard for humans."
His alternative is the centaur benchmark: don't measure the machine alone or the human alone — measure how well they perform together, and optimize for that. It is the same instinct as Klaassen's human sandwich, restated as a design target. His call-center study is the positive-sum proof: customers got better answers, and employees were measurably happier with less turnover. "This can be a win for all three groups if you do it right."
But complementarity has to be designed in. A black-box AI that reads a scan without explaining itself can fail in deployment because it was never built to work with the radiologist. The centaur question — how well does the human-AI team detect the cancer? — forces interpretability, so the human can still catch the eyelash that fell on the slide.
Where they converge
Imas's "interdependent tasks," Shipper's "human sandwich," Solomon's "fewer here, more there," and Brynjolfsson's "centaur" describe the same dynamic from four angles. The unautomated piece isn't a residue — it's the structural reason the rest of the bundle pays off, it's where the human headcount migrates to, and (Brynjolfsson's addition) keeping it human is a design choice, not a given.
04The new scarcity is difference
In a world flooded with cheap AI output, the new scarce good is "made by a human" and "not like everything else." Imas's experiments show the WTP for human-made rises with scarcity. Shipper extends the same logic into the workplace: AI commoditizes yesterday's competence, and the work that doesn't look like everything else becomes the rare, high-status thing.
Willingness to pay — Imas & Madarász experiments
Desire has two parts: how good the thing is (the hedonic), and how few other people can have it (the scarcity). Imas's t-shirt experiment isolated the second piece — and willingness to pay nearly doubled when some buyers were randomly excluded, without changing the product at all.
A follow-up study with Graylyn Mendell ran the same idea on AI-vs-human-made products. When a human-made item is rare, it commands a big premium. When an AI-made item is rare, it doesn't.
That's the engine behind what Imas calls the relational sector. Not just performers — also teachers, doctors, financial advisors, trainers. Their jobs include a piece the customer specifically wants delivered by a human. As AI handles the surrounding tasks, that piece becomes the whole point — and gets more, not less, valuable.
"Because everyone has access to the same models, and the models are all based on yesterday's competence, by default the models end up creating work that ranges from 'a decent start' to 'it's just plain slop.' Slop is not any one particular mistake. Slop is visible sameness, repeated ad nauseam."
"We demand not just any React app or research report — but one that feels exactly right for the person, company, and situation. We demand one that feels alive and specific, not cheap and generic. We want something that has status. When work is abundant and looks alike everywhere, the work that doesn't fit the pattern becomes the rare, valuable, and high-status thing."
"Demand for difference is new demand for experts. The current generation of models only knows about work that has been done. Humans know about what needs to be done, right now, at this moment."
Solomon's list of jobs A.I. can't easily touch is the labor-market mirror of Imas's "relational sector": food preparation, construction, hands-on services. The visible, physical, person-to-person work. "It's hard to see how A.I. is going to do those jobs."
And his cultural exhibits: "TV didn't kill live entertainment. The internet didn't end real estate agents or fitness instructors." The pattern repeats whenever a new technology arrives — the things people specifically want delivered by a person survive, and often appreciate. The premium grows on the human-made side; the automate-able side gets cheaper.
Where they converge
Imas's experimental result, Shipper's "slop" diagnosis, and Solomon's list of resilient hands-on sectors describe the same market. In all three, abundance of the cheap version inflates the value of the rare version — and the rare version is, almost by definition, human.
Imas's prediction: the job titles of 2040 will look familiar — teacher, doctor, planner, engineer. But the day-to-day will invert: AI does most of the bundle, the human delivers the irreplaceable, relational or judgmental piece. And entirely new categories will appear that we can't picture yet — just as most "strongly relational" jobs in 2026 didn't exist in 1940.
05The frame is not the framer
"But what about the benchmarks?" Every doomer's last refuge: AI is climbing exponential curves on Humanity's Last Exam, GDPval, METR. Doesn't that change the story? Shipper's answer — built from inside the lab — is no. Benchmarks measure work inside a frame. Humans choose the frames. That gap is structural, and it's the deepest reason the doom story keeps being wrong.
The benchmarks are real. Humanity's Last Exam: low single digits a year ago, ~44% today. GDPval: from ~lows to ~85% in twelve months. METR's early Claude Mythos result: an 80% success rate on tasks that would take a human four hours.
Shipper calls the temptation to extrapolate from these curves "chart psychosis." The trap: a benchmark is not the territory. To measure anything, you have to freeze a problem into a static — and therefore measurable — frame. Once a frame gets saturated, it is trivial to zero the model out again by changing the frame. Progress continues inside the new frame, of course — but the same process repeats.
The clean illustration is Every's in-house Senior Engineer benchmark, which asks a coding agent to do a first-principles rewrite of a real, vibe-coded codebase. GPT-5.5's best run scored 62/100. Human senior engineers score in the high 80s or low 90s. Next year a model probably catches up. Then the team picks a harder rewrite, asks the model to decide whether a rewrite is needed in the first place, and the gap reopens.
Senior Engineer benchmark, May 2026
The smuggled intelligence. When OpenAI announced GDPval results — "AI as good as experts at half of tasks" — the headlines missed something. Every GDPval task ships with a thousand-word, expert-written prompt that pre-frames the problem. Shipper's line: "There is an enormous amount of human intelligence going into framing this problem in a way that a model can complete." Strip the framing, ask the model "solve all the errors that keep popping up," and the score drops to near zero.
"Even this strong version of AGI does not dissolve the frame problem. This AGI can choose and re-choose frames, but only in pursuit of some goal it has been given, some reward it is optimizing, or some signal someone has decided should count as progress. In any hypothetical AGI built by any of the major labs, there is still going to be a framer — a human — directing the model to achieve a goal."
"Spend 10 minutes with a toddler and you'll see it. A toddler is worse than a language model at almost every task we care about. But the toddler has ends. He invents games constantly. He turns the world into experiments. He is not waiting for a prompt. Current agents have autonomy. They do not have agency."
Solomon's closing rhetorical move plants the same flag from a different direction. In 1930, John Maynard Keynes predicted that by 2030 we'd be working 15-hour weeks thanks to technology. The reality, almost a century later, is closer to 40. The smartest economist of the 20th century missed the shape of the future by an order of magnitude — because he extrapolated inside one frame (output per hour) and missed everything that happened outside it (entirely new categories of work).
"Stark forecasts by even the most brilliant minds often miss their mark." Today's "AI will eliminate half of all white-collar work by 2030" predictions are operating exactly the same way Keynes was — projecting today's frame forward, missing the framers who will keep moving it.
Where this extends Imas — and where Solomon agrees by analogy
Imas's four ideas explain why the labor market keeps absorbing automation. Shipper's fifth explains why it will keep doing so — structurally, even after AGI. Solomon doesn't use the same vocabulary, but his Keynes story is the same insight from a CEO chair: every prior forecast that extrapolated inside one frame got the shape of the future wrong.
06What the numbers actually show
The other three argue mostly from history, theory, and firsthand experience. Brynjolfsson argues from the data — and the data cuts both ways. Most of AI's value is real but invisible to the statistics; the productivity payoff is real but lagged; and the first labor damage is real but narrowly concentrated. Optimism and alarm live in the same dataset. And the one clear labor-market signal is also the most contested number in the literature — which is why, in September, Imas and Jacob Schaal gathered the counter-evidence in one place (below).
The canaries: employment change by AI-exposure, ages 22–25
The upside the statistics can't see. The standard knock on AI is that it isn't showing up in the productivity numbers. Brynjolfsson's reply: a large share of its value was never going to. GDP counts transactions, but much of what AI delivers is free. His team's willingness-to-accept survey asks what you'd need to be paid to give up ChatGPT for a month — about $98 in 2024, $125 in 2025 — value that never crosses a cash register. His proposed fix, GDP-B ("GDP, benefits"), tries to count it: "the average American spends almost half their waking hours looking at screens. They're voting with their time that this free digital stuff is very valuable. It's just not priced."
The J-curve. The second reason the stats lag: a general-purpose technology demands a costly reorganization before it pays off — "to get the full value, you have to rethink how companies organize." That intangible investment can run up to 10× the direct spend on the technology and mostly isn't measured, so output looks flat while the groundwork is laid. His read is that the curve is finally turning: U.S. productivity grew 2.2% in 2025, nearly double the prior decade, with call-center, consulting, and software studies showing double-digit gains.
Canaries Dashboard · annual employment change by age & AI-exposure (through April 2026)
The dashboard is the centerpiece of an expanded partnership between Brynjolfsson's Stanford Digital Economy Lab and ADP Research, whose payroll data covers roughly one in six American workers. ADP chief economist Nela Richardson — Brynjolfsson's partner on the research — argues the key variable is automation vs. augmentation: occupations where AI augments human work show more enduring employment growth; those where AI automates tasks outright contract. Early-career workers, concentrated in the most automatable layer of almost any occupation, sit squarely in the second category.
Her framing mirrors Brynjolfsson's two-questions split. "In the aggregate, AI's impact on jobs remains modest," she wrote in a June 16 post on the first batch of dashboard data — but measured by career stage, "dramatic differences emerge." The headline numbers bear her out: the most AI-exposed occupations contracted just 0.2% year-over-year as of April 2026, versus +0.1% for the least-exposed; since ChatGPT's late-2022 launch, AI-exposed employment still grew 1.1% (against 2% for the least-exposed). At the top line, the sky hasn't fallen. Cut the data by age, and the fault line opens.
The original paper's data ran through July 2025; the dashboard now extends it to June 2026 — nearly four years of post-ChatGPT data. The effect hasn't mean-reverted. It has grown by roughly half a percentage point per month, month after month, and the early-career decline sharpened after year one (−2.8% to April 2024, more than −4% per year by April 2026).
He has stress-tested the finding against the major counter-arguments. The interest-rate hypothesis points the wrong way — the most rate-sensitive occupations, like construction, have the lowest AI exposure. He removed the entire tech sector. He isolated remote-work effects. On his payroll data, the pattern held every time. "If you take out the entire tech industry, or take out all tech-related occupations, or you slice it different ways, you still get this effect." It is the direct rebuttal to the skeptics — among them Jason Furman's "small to zero" reading in the survey below. The dashboard's name is the argument: canaries "didn't stop the danger. They just told you the clock was running."
Two qualifications, from the authors themselves. As Imas and Schaal's September review notes, the within-firm estimates in the latest version of the paper — the specification built to absorb interest-rate and over-hiring shocks — attenuated after an improved data pipeline, and the authors now emphasize the simpler descriptive divergence. The dashboard gap also roughly halves once occupations' education levels are controlled for. The authors present the pattern as descriptive, not causal.
Three vintages of one number · ages 22–25 in the most AI-exposed roles
In September, Imas — writing with Jacob Schaal — reviewed the empirical literature on whether AI has shown up in the labor market yet. They call it a "now-cast," not a forecast, and their verdict is narrower than either camp's headline: "AI may already be affecting the hiring margin for junior white-collar roles most exposed to AI, but this attribution is contested, and aggregate labor-market disruption has yet to appear in the data." The junior-hiring signal replicates across several countries. Population-wide records in the Nordics don't find it. And several studies suggest the decline began before ChatGPT.
| Where | Study · data | What it finds | Reads as |
|---|---|---|---|
| The junior-hiring signal replicates | |||
| U.S. | Brynjolfsson, Chandar & Chen (2025–26) · ADP payroll | Gap for 22–25s in exposed roles widens from 13% (July 2025) to 19% (June 2026) | Signal |
| U.S. | Hosseini & Lichtinger (2026) · 65M workers, 280K firms | Junior employment 8–10% lower at GenAI-adopting firms after two years; seniors flat; driven by slower hiring | Signal |
| U.K. | Klein Teeselink (2025) · Revelio / job postings | Fewer junior positions at exposed firms, emerging ~18 months after ChatGPT | Signal |
| 39 countries | Klein Teeselink & Carey (2026) · online postings | AI exposure linked to a 6.1% drop in postings — larger where employment protection is stricter | Signal |
| Switzerland | Klaeui & Siegenthaler (2025) · unemployment records | Job seekers up 27% in exposed vs. less-exposed occupations (shifts between occupations only) | Signal |
| Sweden | Lodefalk et al. (2026) · employer-employee data | Exposed junior hiring −5.5%, over-50s +1.3% — but a staffing-law change in Oct. 2022 coincides with ChatGPT | Confounded |
| Population-wide records find nothing | |||
| Finland | Kauhanen & Rouvinen (2026) · administrative wages, 2019–25 | No effect of AI exposure on youth wages or employment | Null |
| Norway | Facius & Iacono (2026) · entire population, 2015–Mar. 2025 | No significant effect; exposed junior employment began falling before ChatGPT | Null |
| Denmark | Humlum & Vestergaard (2026) · 25,000 workers, admin records | 93% adoption; precise nulls on earnings, hours and wages — effects above 2% ruled out | Null |
| Timing and remote work | |||
| U.S. | Frank et al. (2026) · UI records, LinkedIn, syllabi | Unemployment risk in exposed occupations rose from early 2022, before ChatGPT | Pre-trend |
| 4 countries | Lambert & Schindler (2026) · job postings, 2017–25 | With remote work in the model, the AI effect "drops to zero or even flips positive" | Remote work |
| U.S. | Brynjolfsson et al. (2026) · same test, ADP payroll | Opposite result: the AI gradient survives and grows; the remote-work gradient shrinks | Signal |
| Measured by adoption, not exposure | |||
| U.S. | Kharazian, Simon & Stevens (2026) · AI spend, 21,000 firms | High-intensity adopters grow employment 10% and entry-level employment 12%; sample not representative | Growth |
Findings as summarized by Imas & Schaal. Their caveat on the Nordic nulls: population records capture employment levels, not the faster-moving hiring flows where the U.S. and U.K. signal shows up.
The sharpest challenge is Lambert and Schindler's. Remote work and AI exposure cluster in the same white-collar, computer-heavy jobs, and remote work on its own raises the cost of supervising juniors and cuts what they learn by watching seniors. Put both in the same model and the remote-work effect barely moves while the AI effect disappears. Brynjolfsson's team reran the identical specification on ADP payroll data and got the reverse. Because the test is the same, they attribute the difference to the data (job postings and LinkedIn profiles versus payroll records), and possibly to measuring seniority by role rather than by age. Hosseini and Lichtinger, meanwhile, find the junior decline at AI-adopting firms just as pronounced in jobs that can't be done remotely.
Imas and Schaal's read is that the two causes may not be rivals. They borrow an image from David Deming — the same Deming who appears in the WSJ survey below: remote work "containerized" office work. It made whole jobs fully digital, with a data trail and no stigma about using AI, which is exactly where AI could then be deployed. Remote work weakened supervision; AI slotted into the workflows it left behind.
Outside that junior margin, the lagging indicators have barely moved. Executives cite AI as a major reason for layoffs, but only 1% of laid-off workers attribute their firing to AI (Gallup) — the worker's-eye version of the point at the top of this card that AI is the headline, not the cause. A survey of nearly 6,000 executives in the U.S., U.K., Germany and Australia (Yotzov et al.) finds 69% of firms using AI and about 90% saying it hasn't significantly affected employment or productivity so far. Executives do expect employment reductions over the next three years — and in a U.K. subsample, about two-thirds of the anticipated cut comes from hiring less rather than letting people go. Google's ATLAS usage data suggests why the totals are calm: AI touches jobs covering 88% of U.S. employment, but only 21% of the tasks within them. Broad, shallow adoption — not yet the automation of whole workflows.
Their closing caution cuts both ways. Agentic AI in enterprise workflows "is in its infancy," and as bottlenecks break, the junior-hiring pattern "may then begin to radiate through the rest of the labor market." But messy jobs whose tasks resist unbundling (the O-ring logic of Concept 03), existing and new regulation, and new relational and startup work all push the other way.
The reason the data can look contradictory is that it answers two different questions. The aggregate question — will there be more productivity, more value, probably more jobs eventually? — he answers with real confidence. The distributional question — who captures the gains, and how fast can the displaced adjust? — he refuses to wave away. Most optimists and most doomers collapse the two into one verdict. Brynjolfsson keeps them apart, and it is the single move that most separates him from the other three voices here: they are largely arguing the first question; his deepest worry lives in the second. By September the gap had narrowed from Imas's side: his review with Jacob Schaal lands on the same split, in nearly the same words — "aggregate stability currently coexists with subgroup displacement."
07Sixteen economists, one ballot
The four voices above argue. In June 2026, The Wall Street Journal put the same questions to a vote — surveying sixteen economists, including 2024 Nobel laureate Daron Acemoglu, three former chairs and advisers of the White House Council of Economic Advisers, and the academics whose research frames this entire document (Autor's task model and Gans's O-ring underpin Concept 03; here they answer for themselves). The result is the best reality check in this compilation: unanimity on productivity, a genuine split on jobs — and the split is a coherent bloc, not noise.
The ballot — five questions, sixteen economists
The survey, read against the six lenses
Where the sixteen land on the arguments above — supporters and dissenters both. Each chip links back to the matching concept.
Five of the sixteen — Acemoglu, Chandar, Henderson, Restrepo and Wolfers — vote both "net job losses" and "replace." This is not a fringe: it includes a Nobel laureate whose life's work is automation. Notice, though, what the bloc disputes. Restrepo states the scarcity principle of Concept 04; Acemoglu endorses the interpersonal turn; Wolfers gives framer advice. They mostly accept the mechanisms of Concepts 01–05 — what they dispute is the pace, the geography, and whether institutions can absorb the shock in time. That is precisely the line Brynjolfsson draws in Concept 06, which makes the survey, on balance, a sixteen-economist vote for keeping the aggregate and distributional questions separate.
Who the five sources draw on
Imas works from a 200-year economic tradition. Shipper works from the last three years inside Every. Solomon argues from Goldman's headcount across six decades. Brynjolfsson brings the payroll data and a measurement lab. Here's the cast all five draw on — note that Autor and Gans appear twice, as cited theory and as WSJ survey respondents.
What the five agree on — and don't
Step back from the individual ideas and the five sources resolve into a single shape: a broad agreement about the aggregate, and four different anxieties about the texture. They converge on what automation does to the quantity of work. They part ways on whether the resulting world is one we can actually live in — and Brynjolfsson is the one who insists those are two different questions that deserve two different answers. The WSJ ballot turns the shape into a head count: unanimous on productivity, eight-to-five for complement over replace — a majority verdict carrying a Nobel-led minority report.
Where the five agree
Where they part ways — four anxieties beneath one optimism
Prehistoric brains. Medieval institutions. Godlike technology.
Brynjolfsson's deepest worry isn't too few jobs — it's that the gap between what AI can already do and what our institutions can absorb produces severe disruption before we adapt. Even his optimistic story still requires a transition the labor market isn't built to run at speed: labor mobility has fallen over 20–30 years, universities aren't teaching AI-native work, and the BLS is being cut "at the exact wrong time." The capital/labor ratio, roughly constant for 200 years, has begun to drop — and a smaller wage share means a smaller share of people with economic, and ultimately political, power.
His prescription isn't to slow the technology but to invest in absorbing it: better measurement (GDP-B, the Stanford indicators), more dynamic labor markets, and centaur benchmarks that steer AI toward complementing people. The aggregate gains, he argues, are earned. The widely-shared prosperity is not — it has to be chosen.
Optimism is conditional, not complacent.
Solomon's case is a rebuttal built on one asymmetry: every previous automation wave triggered an apocalypse forecast that turned out wrong — Keynes among the forecasters — so the burden of proof sits with whoever claims this time is different. But he names the real risk openly: this transition could move faster than the reskilling system can absorb, and the U.S. has let its training infrastructure lapse before.
The posture is humility, not panic: resilience as the prior, apocalypse as the burden of proof — and a reskilling bill that comes due whether or not the optimists are right. It is the policy-flavored echo of Brynjolfsson's institutional gap.
If the relational sector grows, who's left to do the relating?
Imas's optimism rests on a deep human desire for connection — the barista who knows your name, the teacher who reads the room, the doctor who sits with you. As AI scales, the dollar value of delivered-by-human work should rise.
And yet: time spent socializing is falling, indoor hours are rising, and a 2026 paper in Psychological Science found that loneliness increased the more people leaned on chatbot companions.
A humming relational economy that monetizes our base desire for connection — even as the unpriced, ordinary kind keeps disappearing. The economy may be fine. The loneliness may not be.
I found my clothes — but where am I myself?
Shipper closes with a story from Martin Buber's The Way of Man. A man, exhausted by losing his clothes every morning, wrote down where he'd put each item. The next morning he read the slip — "cap," "pants" — and dressed perfectly. Then:
The agents find your clothes — your code, your inbox, your draft slide deck. What they can't do, because they have no ends of their own, is locate you. The frame is not the framer. The job titles of 2040 may look familiar; who is doing the framing inside them is the question no benchmark can answer for you.